This report describes a 'learning program' that acquires much of the knowledge required by a parsing system that processes conversations in a 'natural' language akin to ham-radio jargon. The learning program derives information from example sentences taken from transcripts of actual conversations, and uses this knowledge to extend the 'core' augmented transition network (ATN) grammar. The parser can use the extended grammar to process the example sentences, plus a large number of syntactically and semantically related sentences. The learning program uses a set of heuristics to determine the difference between the existing version of the grammar and a superset that could process the example sentence. A set of models act as templates to produce possible extensions to the grammar. An evaluation measure selects one of the extensions and adds it to the grammar. This extension is henceforth an integral component of the knowledge base and may be used by the parser to process conversations and by the learning program to extend the grammar further. This report relates the mechanisms used by the learning program to grammatical inference of context-sensitive languages, which include the natural languages, and some proposed linguistic models of human language acquisition. These models describe language acquisition as a process of developing hypotheses according to the constraints of innate universal rules, and acceptance of those hypotheses that make it possible for the child to understand new sentences.

plus-circle Add Review

comment

Reviews